Using machine learning to improve national lake depth predictions
2024-11-19
FENZ Lakes
Depth contours
Bathymetry raster
Estimated lake bed
Estimated hypsographic curve
Hypsographic summary
| Variable | Value |
|---|---|
| Max depth | 90.40 |
| Mean depth | 45.33 |
| Volume development index | 1.50 |
| Area (ha) | 1063.51 |
| Shoreline length (km) | 23.40 |
| Shoreline development index | 2.04 |
Lakes with bathymetric data are relatively evenly distributed across New Zealand.
Lake geomorphic type is highly regionalised.
Lakes which have maximum depth data but no bathymetry data highlight that there is probably a further bathymetry data out there.
Some lakes, particularly glacial lakes might have quite a dynamic bathymetry.
Development of Volume, \(D_V\) (Hutchinson 1957) is a measure of departure of the shape of the lake basin from that of a cone calculated using the maximum depth \(Z_{max}\) and the average depth \(\bar{Z}\) :
\[D_V = \frac{3 \times \bar{Z}}{Z_{max}}\] \(D_V\) is greatest in shallow lakes with flat bottoms and low in deep lakes with steep sides.


| Source | N | Bias | RMSE | R2 |
|---|---|---|---|---|
| FENZ | 183 | -0.79 | 44.64 | 0.70 |
| GloBATHY | 119 | 20.93 | 72.68 | 0.43 |
| ML model | 190 | -0.49 | 19.17 | 0.95 |
| Source | N | Bias | RMSE | R2 |
|---|---|---|---|---|
| FENZ | 155 | 6.24 | 21.59 | 0.71 |
| Hydrolakes | 91 | 7.27 | 22.48 | 0.79 |
| ML model | 155 | -0.15 | 7.80 | 0.97 |
Bathymetric data from 156 lakes were digitised and collated for use in this study.
Using a machine learning model, we were able to predict maximum lake depth from morphometric data with an \(r^2\) = 0.95 and mean lake depth with an \(r^2\) = 0.97.
This was substantially better than the FENZ dataset, which had an \(r^2\) = 0.7 for max depth and 0.71 for mean depth.
This project was funded by the New Zealand Ministry of Business, Innovation and Employment (MBIE) through the Smart Ideas program within the Endeavour Fund (grant UOWX2103).
We would like to thank all our data contributors who provided data for the LERNZmp project.
